Is there a way of testing the performance of SSE / AVX / FMA etc. on an individual system?
Once there was a benchmark package available, but I could not find it anymore and it was outdated (more than 10 years) too.

Does anyone know a way how to test for the best instruction set apart from letting boinc choose based on results after some time.

Because I noticed other important metrics too, like system responsiveness or power consumption vs. speed or in other words efficiency.

you can test each application individually
https://asteroidsathome.net/boinc/download/period_search_10214_windows_x86_64__sse3_win10.exe
https://asteroidsathome.net/boinc/download/period_search_10213_windows_x86_64__avx_win10.exe
https://asteroidsathome.net/boinc/download/period_search_10213_windows_x86_64__fma_win10.exe

download a sample workunit, rename it to "period_search_in" and place it in the same folder. Then you can reliably compare runtimes for each version.

results may vary based on the CPU you have, there are also other things to consider - power consumption, running multiple WUs, operating system etc.

usually, AVX/FMA apps have similar performance. Some server Xeons have better FMA performance so they may run faster with FMA. On the other hand, even if Bulldozer/Piledriver support superior AVX, it's actually slower than SSE3.

to give you a rough idea about general performance, here are (single thread) results on some architectures I have